Notes from my August liked tweets and GitHub stars. Lot of good numbers came out this month on what happens when you scale internal coding/ops agents past the prototype stage - especially around skill routing limits and KV cache bills.
Main takeaways from August
- Tool and skill selection falls apart past ~150 skills. Stripe’s internal Kai platform (built on Deep Agents) hit 83% weekly adoption for generating reports, dashboards, and docs [1]. Their biggest practical finding: once you dump more than ~150 skills/tools into an agent’s catalog, selection accuracy drops off a cliff. You gotta hierarchical-route to the domain first, then load only those tools.
- Plain CLIs still beat MCP schemas by 5-28x on tokens. Both Kimi Developers [2] and Pi [3] ran benchmarks showing that for standard dev workflows, letting the model call CLI tools (
gh,git,jq,curl) matches MCP reliability at 5-28x lower token usage. - Switching models mid-session busts your KV prefix cache. It sounds clever to route easy turns to Model A and hard turns to Model B inside the same conversation, until you realize every provider switch throws away your cached prompt prefix and charges you full prefill on a 100k-token history [4] [5]. Artificial Analysis even had to publish a corrected leaderboard coz cache-hit pricing completely flips which models are cheapest on real multi-turn agent tasks [6]. If you use a cheap model, spawn it in an isolated subagent with a tiny context window.
- Pin exact model snapshots, not provider aliases.
@satyajeetjadhavshared a painful story where an API alias kept the exact same string name while swapping the underlying model and billing tier underneath, quietly adding ₹6 lakh to their bill [7]. Always pin dated model IDs in production. - Retained reasoning + compaction is a massive lever. OpenAI’s “Codex as a platform” writeup showed that keeping reasoning state across compactions jumped ARC-AGI-3 from 13.3% to 38.3% while cutting output tokens by 6x.
Software factories & codebase loops
- Uber’s software factory: 70%+ of PRs at Uber now come through agents backed by 3,600 skills and a 24M-node context graph. Coolest detail is how they handled a 9x jump in usage at flat spend - they defaulted subagents to smaller models and tuned their prompt cache TTLs so multi-step loops stayed inside the cache window.
- DHH’s Python-to-Rust port with planner + executor split: DHH ported a Python terminal-effects library to Rust -> startup dropped from 87ms to 2ms, rendering got 9.6x faster, zero runtime dependencies, single 3MB binary [8]. When he took the frontier model’s architecture plan and handed execution to a cheaper model, the port finished in 84 minutes for ~$55 [9] (and DeepSeek Pro knocked out a similar run in 2.5 hours for $23 vs $550 on Fable, while smaller models stalled out [10]).
- Boris Cherny’s daily dead-code & fuzzing agents: running scheduled background agents that fuzz iOS/Android simulators for crashes, unify duplicate helpers, and add runtime logging to suspected dead code paths before opening deletion PRs -> 388 PRs opened, 180 merged. [11]
Agent infra under the hood
- Cursor’s WAL-backed Git for cloud agents: spinning up thousands of short-lived cloud agent sandboxes melted traditional 3-replica Git servers. Cursor replaced them with an object-storage write-ahead log (WAL) + local NVMe cache and pre-warmed environments, cutting cloud agent startup time by 3x. [12]
- Headlong (
shellm): Andy Konwinski demoed a <10k-line Bash microharness running a continuous 24/7 JSONL DAG thought stream at $1-2/hr that diagnosed and fixed an unwired memory recall bug on its own in 48 minutes. [13]
Starred GitHub repos from August
- herdrdev/herdr - Rust terminal multiplexer built specifically for coding agents (
herdr). Tracks which panes are working, blocked on permission, or idle across local + SSH machines, and exposes a Unix socket API so agents can spawn and message sibling panes. - deepseek-ai/deepseek-harness & awesome-dsh-plugin - DeepSeek’s open-source agent harness (
dsh) where providers, tools, compaction strategies, and TUI panes are all hot-swappable plugins - nexu-io/open-design - local-first open-source alternative to Claude Design with 151
DESIGN.mdbrand systems, HTML-to-MP4 video rendering, and direct handoff to 25+ coding agent CLIs - cloudflare/computer - SQLite-backed virtual filesystem inside a Cloudflare Durable Object that mounts into FUSE containers,
just-bashisolate shells, or JS workers - kepano/obsidian-skills - Steph Ango’s official
SKILL.mdpack for Obsidian Markdown, Bases (.base), JSON Canvas, and clean web extraction via Defuddle - dragthelake/ambient-context - lightweight Rust macOS menu bar app that captures the accessibility tree of your focused window into deduplicated daily markdown files so your coding agent has context on docs/threads you read earlier that day
- microsoft/agent-governance-toolkit - deterministic policy middleware and kill-switches for gating agent tool calls outside the prompt
- milind-soni/OpenMausBot & b-nnett/grok-bot-0.18-reconstructed - open-source multi-agent workspace and VM desktop setups inspired by Grok Bot